double-agent（双重间谍）

DoubleAgent Skill

Purpose

The DoubleAgent pattern solves a fundamental problem in AI-generated software: AI self-evaluation bias.

When a single AI agent both generates and evaluates its own output, it systematically overestimates quality — the same cognitive conflict that occurs when a student grades their own exam. The solution is to forcibly separate the two cognitive roles into independent agents with different prompts, goals, and evaluation criteria.

This skill provides:

1. Architecture templates for Generator-Evaluator agent pairs
2. Evaluator prompt templates calibrated with few-shot scoring examples
3. Iteration loop design for 5-15 round refinement cycles
4. Playwright integration patterns for real browser-based evaluation
5. Scoring rubric design to prevent score drift and grade inflation

Core Architecture

User Goal / Spec
      ↓
 ┌─────────────┐
 │  Generator  │ ← Produces output (code, UI, content, data)
 └──────┬──────┘
        │ output artifact
        ↓
 ┌────────────────────────────────────┐
 │           Evaluator                │
 │  • Reads spec (NOT generator output)│
 │  • Operates artifact via Playwright │
 │    (click, fill form, navigate)     │
 │  • Scores on rubric (0-100)         │
 │  • Writes structured feedback       │
 └────────────────┬───────────────────┘
                  │ score + feedback
                  ↓
         ┌────────────────┐
         │ Score ≥ target? │
         │   YES → Done    │
         │   NO → Loop     │
         └────────┬────────┘
                  │
                  └──→ Generator (next iteration)

Key principle: The Evaluator reads the original spec, not the Generator's output. It evaluates independently, as if it were a real user encountering the product for the first time.

When to Apply

Implementation Steps

Step 1: Define the Spec Contract

Write a clear spec that both agents will reference independently. The spec must be:

• Concrete (measurable outcomes, not vague goals)
• Observable (evaluable through interaction or inspection)
• Versioned (so both agents work from the same contract)

See references/architecture.md for spec template.

Step 2: Configure the Generator Agent

Assign the Generator a single role: produce output that satisfies the spec.

• Do NOT ask the Generator to self-evaluate
• Do NOT include evaluation criteria in the Generator's prompt
• Provide: spec + iteration history + previous evaluator feedback

Step 3: Configure the Evaluator Agent

Assign the Evaluator a single role: independently verify the spec is satisfied.

• Load references/evaluator-prompts.md for calibrated prompt templates
• Use Playwright MCP for UI/web artifacts (real browser interaction)
• Use structured JSON output for scores to enable automated loop control
• Calibrate with few-shot examples BEFORE running (prevents grade inflation)

Step 4: Design the Iteration Loop

MAX_ROUNDS = 15
PASS_THRESHOLD = 80  # out of 100

for round in range(MAX_ROUNDS):
    output = generator.run(spec, history)
    evaluation = evaluator.run(spec, output)  # Playwright-based
    
    history.append({"round": round, "score": evaluation.score, "feedback": evaluation.feedback})
    
    if evaluation.score >= PASS_THRESHOLD:
        break
    
    if evaluation.score_trend == "plateauing":
        generator.switch_approach()  # Complete strategy reset

See scripts/iteration_loop.py for a complete implementation template.

Step 5: Calibrate the Evaluator

To prevent score drift, run the Evaluator on 3-5 known examples FIRST:

• 1 example at ~30/100 (clearly bad)
• 1 example at ~60/100 (mediocre)
• 1 example at ~85/100 (good)
• 1 example at ~95/100 (excellent)

If scores deviate >15 points from expected, adjust the Evaluator's prompt or rubric weights before the real run.

Scoring Rubric Design

Effective rubrics for software systems:

Adjust weights based on the domain. For content systems, increase "originality". For data pipelines, increase "edge case handling".

Playwright Integration (for UI artifacts)

When evaluating web/H5/mini-program outputs, the Evaluator should:

1. Navigate to the deployed artifact URL
2. Execute each spec requirement as a user action sequence
3. Observe actual behavior (DOM state, network requests, visual output)
4. Record pass/fail per requirement with screenshots
5. Report structured JSON with score breakdown

Playwright MCP tool calls to use:

• playwright_navigate → open URL
• playwright_click → interact with elements
• playwright_fill → fill form inputs
• playwright_screenshot → capture evidence
• playwright_get_visible_text → verify content

Reference Files

• references/architecture.md — Detailed architecture patterns, spec templates, and design rationale
• references/evaluator-prompts.md — Ready-to-use Evaluator prompt templates for different artifact types

Scripts

• scripts/iteration_loop.py — Complete iteration loop implementation template
• scripts/calibrate_evaluator.py — Evaluator calibration utility